Learning-Based Compositional Parameter Synthesis for Event-Recording Automata - Inria - Institut national de recherche en sciences et technologies du numérique Access content directly
Conference Papers Year : 2017

Learning-Based Compositional Parameter Synthesis for Event-Recording Automata

Abstract

We address the verification of timed concurrent systems with unknown or uncertain constants considered as parameters. First, we introduce parametric event-recording automata (PERAs), as a new subclass of parametric timed automata (PTAs). Although in the non-parametric setting event-recording automata yield better decidability results than timed automata, we show that the most common decision problem remains undecidable for PERAs. Then, given one set of components with parameters and one without, we propose a method to compute an abstraction of the non-parametric set of components, so as to improve the verification of reachability properties in the full (parametric) system. We also show that our method can be extended to general PTAs. We implemented our method, which shows promising results.
Fichier principal
Vignette du fichier
446833_1_En_2_Chapter.pdf (417.23 Ko) Télécharger le fichier
Origin : Files produced by the author(s)
Loading...

Dates and versions

hal-01658415 , version 1 (07-12-2017)

Licence

Attribution

Identifiers

Cite

Etienne André, Shang-Wei Lin. Learning-Based Compositional Parameter Synthesis for Event-Recording Automata. 37th International Conference on Formal Techniques for Distributed Objects, Components, and Systems (FORTE), Jun 2017, Neuchâtel, Switzerland. pp.17-32, ⟨10.1007/978-3-319-60225-7_2⟩. ⟨hal-01658415⟩
132 View
93 Download

Altmetric

Share

Gmail Facebook X LinkedIn More